This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: match.pd: Optimize (x & y) ^ (x | y)

On Thu, Jun 11, 2015 at 05:25:30PM +0200, Marc Glisse wrote:
> On Thu, 11 Jun 2015, Marek Polacek wrote:
> >I have verified this transformation on a toy testcase (tried x and y
> >in the range [-1000,1000]) and it does a correct thing for all integers.
> Note that for pure bitop (only involving &|^), testing the range [0,1] is
> sufficient.
I'd feel safer when testing a wider range of integers ;).

> >+/* (x & y) ^ (x | y) -> x ^ y */
> >+(simplify
> >+ (bit_xor:c (bit_and@2 @0 @1) (bit_ior@3 @0 @1))
> Make either bit_and or bit_ior commutative? Or do we canonicalize in a way
> that makes it unnecessary?

Correct: bit_and and bit_ior don't need :c here.  (But the bit_xor needs it.)
I've tried various ordering of x and y and all of them were optimized.
Arguably I should've put more tests into the testcase...

> >+  (if (single_use (@2) && single_use (@3))
> >+   (bit_xor @0 @1)))
> I don't think we should use single_use here. The result is never more
> complicated than the original. Sure, it might increase register pressure a
> bit in some cases, but we have not used that as a criterion for other
> simplifications in match.pd yet (LLVM does though).

I don't have a strong preference here but we surely use single_use
in match.pd elsewhere.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]